[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [Qemu-arm] [Qemu-devel] [PATCH for 4.1] aspeed/timer: Provide back-p
From: |
Philippe Mathieu-Daudé |
Subject: |
Re: [Qemu-arm] [Qemu-devel] [PATCH for 4.1] aspeed/timer: Provide back-pressure information for short periods |
Date: |
Thu, 4 Jul 2019 11:26:53 +0200 |
User-agent: |
Mozilla/5.0 (X11; Linux x86_64; rv:60.0) Gecko/20100101 Thunderbird/60.7.0 |
CC'ing Stefan & Paolo for a non-ARM view on this...
On 7/4/19 7:51 AM, Cédric Le Goater wrote:
> From: Andrew Jeffery <address@hidden>
>
> First up: This is not the way the hardware behaves.
>
> However, it helps resolve real-world problems with short periods being
> used under Linux. Commit 4451d3f59f2a ("clocksource/drivers/fttmr010:
> Fix set_next_event handler") in Linux fixed the timer driver to
> correctly schedule the next event for the Aspeed controller, and in
> combination with 5daa8212c08e ("ARM: dts: aspeed: Describe random number
> device") Linux will now set a timer with a period as low as 1us.
>
> Configuring a qemu timer with such a short period results in spending
> time handling the interrupt in the model rather than executing guest
> code, leading to noticeable "sticky" behaviour in the guest.
>
> The behaviour of Linux is correct with respect to the hardware, so we
> need to improve our handling under emulation. The approach chosen is to
> provide back-pressure information by calculating an acceptable minimum
> number of ticks to be set on the model. Under Linux an additional read
> is added in the timer configuration path to detect back-pressure, which
> will never occur on hardware. However if back-pressure is observed, the
> driver alerts the clock event subsystem, which then performs its own
> next event dilation via a config option - d1748302f70b ("clockevents:
> Make minimum delay adjustments configurable")
>
> A minimum period of 5us was experimentally determined on a Lenovo
> T480s, which I've increased to 20us for "safety".
>
> Signed-off-by: Andrew Jeffery <address@hidden>
> [clg: - changed the computation of min_ticks to be done each time the
> timer value is reloaded. It removes the ordering issue of the
> timer and scu reset handlers but is slightly slower ]
> - introduced TIMER_MIN_NS
> - introduced calculate_min_ticks() ]
> Signed-off-by: Cédric Le Goater <address@hidden>
> ---
>
> This is fixing a serious slowdown issue with recent Linux.
>
> hw/timer/aspeed_timer.c | 17 ++++++++++++++++-
> 1 file changed, 16 insertions(+), 1 deletion(-)
>
> diff --git a/hw/timer/aspeed_timer.c b/hw/timer/aspeed_timer.c
> index 29cc5e807081..217d59fa7885 100644
> --- a/hw/timer/aspeed_timer.c
> +++ b/hw/timer/aspeed_timer.c
> @@ -42,6 +42,13 @@ enum timer_ctrl_op {
> op_pulse_enable
> };
>
> +/*
> + * Minimum value of the reload register to filter out short period
> + * timers which have a noticeable impact in emulation. 5us should be
> + * enough, use 20us for "safety".
> + */
> +#define TIMER_MIN_NS (20 * SCALE_US)
> +
> /**
> * Avoid mutual references between AspeedTimerCtrlState and AspeedTimer
> * structs, as it's a waste of memory. The ptimer BH callback needs to know
> @@ -96,6 +103,14 @@ static inline uint32_t calculate_ticks(struct AspeedTimer
> *t, uint64_t now_ns)
> return t->reload - MIN(t->reload, ticks);
> }
>
> +static uint32_t calculate_min_ticks(AspeedTimer *t, uint32_t value)
> +{
> + uint32_t rate = calculate_rate(t);
> + uint32_t min_ticks = muldiv64(TIMER_MIN_NS, rate,
> NANOSECONDS_PER_SECOND);
> +
> + return value < min_ticks ? min_ticks : value;
> +}
> +
> static inline uint64_t calculate_time(struct AspeedTimer *t, uint32_t ticks)
> {
> uint64_t delta_ns;
> @@ -259,7 +274,7 @@ static void aspeed_timer_set_value(AspeedTimerCtrlState
> *s, int timer, int reg,
> switch (reg) {
> case TIMER_REG_RELOAD:
> old_reload = t->reload;
> - t->reload = value;
> + t->reload = calculate_min_ticks(t, value);
>
> /* If the reload value was not previously set, or zero, and
> * the current value is valid, try to start the timer if it is
>